• ColonelThirtyTwo@pawb.social
    link
    fedilink
    arrow-up
    24
    arrow-down
    5
    ·
    3 months ago

    SQL blows for hierarchical data though.

    Want to fetch a page of posts AND their tags in normalized SQL? Either do a left join and repeat all the post values for every tag or do two round-trip queries and manually join them in code.

    If you have the tags in a JSON blob on the post object, you just fetch and decide that.

    • blackstrat@lemmy.fwgx.uk
      link
      fedilink
      arrow-up
      7
      ·
      3 months ago

      I’m no expert in JSON, but don’t you lose the ability to filter it before your application receives it all? If you had a reasonable amount of data then in SQL you can add WHERE clause and cut down what you get back so you could end up processing a lot less data than in your JSON example, even with the duplicated top table data. Plus if you’re sensible you can ensure you’re not bringing back more fields than you need.

      • thewebroach@lemmy.world
        link
        fedilink
        arrow-up
        2
        ·
        edit-2
        3 months ago

        If there’s commonly used data that would be good for indexing or filtering, you can take a few key values and keep them stored in their own fields.

        There are also often functions that can parse structured text like XML or JSON so you can store data in blobs but not actually need to query all the blobs out to a client to use them on the database side and retrieve specific values. Another nice thing about blobs is the data can be somewhat flexible in structure. If i need to add a field to something that is a key/value pair inside a blob, i dont necessarily have to change a bunch of table schemas to get the functionality on the front end that I’m after. Just add a few keys inside the blob.

      • Ephera@lemmy.ml
        link
        fedilink
        arrow-up
        2
        ·
        3 months ago

        In a traditional SQL database, yeah. In various document-oriented (NoSQL) databases, though, you can do that.

    • Vlyn@lemmy.zip
      link
      fedilink
      arrow-up
      4
      ·
      3 months ago

      If you only join on indexed columns and filter it down to a reasonable number of results it’s easily fast enough.

      For true hierarchical structures there’s tricks. Like using an extra Path table, which consists of AncestorId, DescendentId and NumLevel.

      If you have this structure:

      A -> B -> C

      Then you have:

      A, A, 0

      A, B, 1

      A, C, 2

      B, B, 0

      B, C, 1

      C, C, 0

      That way you can easily find out all children below a node without any joins in simple queries.

      • ColonelThirtyTwo@pawb.social
        link
        fedilink
        arrow-up
        2
        ·
        3 months ago

        The fact that you’d need to keep this structure in SQL and make sure it’s consistent and updated kinda proves my point.

        It’s also not really relevant to my example, which involves a single level parent-child relationship of completely different models (posts and tags).

        • Vlyn@lemmy.zip
          link
          fedilink
          English
          arrow-up
          1
          ·
          3 months ago

          I mean in my case it’s for an international company where customers use this structure and the depth can basically be limitless. So trying to find the topmost parent of a child or getting all children and their children anywhere inside this structure becomes a performance bottleneck.

          If you have a single level I really don’t understand the problem. SQL joins aren’t slow at all (as long as you don’t do anything stupid, or you start joining a table with a billion entries with another table with a billion entries without filtering it down to a smaller data subset).

          • ColonelThirtyTwo@pawb.social
            link
            fedilink
            arrow-up
            1
            ·
            edit-2
            3 months ago

            My point is that SQL works with and returns data as a flat table, which is ill fitting for most websites, which involve many parent-child object relationships. It requires extra queries to fetch one-to-many relationships and postprocessing of the result set to match the parents to the children.

            I’m just sad that in the decades that SQL has been around, there hasn’t been anything else to replace it. Most NoSQL databases throw out the good (ACID, transactions, indexes) with the bad.

            • Vlyn@lemmy.zip
              link
              fedilink
              English
              arrow-up
              2
              ·
              3 months ago

              I really don’t see the issue there, you’re only outputting highly specific data to a website, not dumping half the database.

              Do you mean your typical CRUD structure? Like having a User object (AuthId, email, name, phone, …), the user has a Location (Country, zip, street, house number, …), possibly Roles or Permissions, related data and so on?

              SQL handles those like a breeze and doesn’t care at all about having to resolve the User object to half a dozen other tables (it’s just a 1…1 relation, on 1…n, but with a foreign key on the user id it’s all indexed anyway). You also don’t just grab all this data, join it and throw it to the website (or rather the enduser API), you map the data to objects again (JSON in the end).

              What does it matter there if you fetched the data from a NoSQL document or from a relational database?

              The only thing SQL is not good at is if you have constantly changing fields. Then JSON in SQL or NoSQL makes more sense as you work with documents. For example if you offer the option to create user forms and save form entries. The rigid structure of SQL wouldn’t work for a dynamic use-case like that.

    • ShortFuse@lemmy.world
      link
      fedilink
      arrow-up
      2
      ·
      edit-2
      3 months ago

      Either do a left join and repeat all the post values for every tag or do two round-trip queries and manually join them in code.

      JSON_ARRAYAGG. You’ll get the object all tidied up by database in one trip with no need to manipulate on the receiving client.

      I recently tried MariaDB for a project and it was kinda neat, having only really messed with DynamoDB and 2012 era MsSQL. All the modern SQL languages support it, though MariaDB and MySQL don’t exactly follow the spec.